Application of large language models for analysing the malicious nature of text

Large language models have now achieved impressive results across a wide range of natural language processing tasks, demonstrating high performance with zero or limited training data. Tasks such as sentiment analysis and named entity recognition have particularly benefited from these advances, highlighting the potential of models to perform tasks without extensive task-specific data sets. Currently, research has focused on the application of large language models in text moderation, which plays an important role in managing user content in various formats, including the detection of malicious and toxic text in social media, online chat messages, forum discussions, website comments, etc. This article examines the features of using various large language models to detect harmful text and presents the results of their comparison with both similar models and other approaches to detecting hostile content.